[RFT PATCH v2 2/4] arm64: restore FPSIMD to default state for kernel and signal contexts

[will.deacon@arm.com (Will Deacon)]

On Mon, Oct 14, 2013 at 04:30:00PM +0100, Jiang Liu wrote:
> On 10/14/2013 11:16 PM, Will Deacon wrote:
> > On Sun, Oct 13, 2013 at 03:20:18PM +0100, Jiang Liu wrote:
> >> From: Jiang Liu <jiang.liu@huawei.com>
> >>
> >> Restore FPSIMD control and status registers to default values
> >> when creating new FPSIMD contexts for kernel context and reset
> >> FPSIMD status register when creating FPSIMD context for signal
> >> handling, otherwise the stale value in FPSIMD control and status
> >> registers may affect the new kernal or signal handling contexts.
> >>
> >> Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
> >> Cc: Jiang Liu <liuj97@gmail.com>
> >> ---
> >>  arch/arm64/include/asm/fpsimd.h | 16 ++++++++++++++++
> >>  arch/arm64/kernel/fpsimd.c      | 11 +++++++++--
> >>  arch/arm64/kernel/signal.c      |  1 +
> >>  arch/arm64/kernel/signal32.c    |  1 +
> >>  4 files changed, 27 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/arch/arm64/include/asm/fpsimd.h b/arch/arm64/include/asm/fpsimd.h
> >> index c43b4ac..b2dc30f 100644
> >> --- a/arch/arm64/include/asm/fpsimd.h
> >> +++ b/arch/arm64/include/asm/fpsimd.h
> >> @@ -50,8 +50,24 @@ struct fpsimd_state {
> >>  #define VFP_STATE_SIZE		((32 * 8) + 4)
> >>  #endif
> >>  
> >> +#define	AARCH64_FPCR_DEFAULT_VAL	0
> >> +
> >>  struct task_struct;
> >>  
> >> +static inline void fpsimd_init_hw_state(void)
> >> +{
> >> +	int val = AARCH64_FPCR_DEFAULT_VAL;
> >> +
> >> +	asm ("msr fpcr, %x0\n"
> >> +	     "msr fpsr, xzr\n"
> >> +             : : "r"(val));
> >> +}
> >> +
> >> +static inline void fpsimd_clear_fpsr(void)
> >> +{
> >> +	asm ("msr fpsr, xzr\n");
> >> +}
> > 
> > You have pretty weak asm constraints here...
> Hi Will,
> 	We will add an explicit "volatile" here. But according to GCC docs, it
> should have the same effect:
> An asm instruction without any output operands is treated identically to
> a volatile asm instruction.

I don't think volatile is enough to prevent re-ordering across a function
call; it just prevents the block from being optimised away entirely and/or
reordered with respect to other volatile statements.

A "memory" clobber should do the trick in this case.

Will